forked from shibing624/MedicalGPT
-
Notifications
You must be signed in to change notification settings - Fork 0
/
supervised_finetuning.py
940 lines (862 loc) · 35.7 KB
/
supervised_finetuning.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
# -*- coding: utf-8 -*-
# Copyright 2023 XuMing(xuming624@qq.com) and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""
Fine-tuning the library models for causal language modeling (GPT, GPT-2, CTRL, ...) on a text file or a dataset.
part of this code is adapted from https://github.com/shibing624/textgen
"""
import math
import os
from dataclasses import dataclass, field
from glob import glob
from typing import List, Sequence, Optional, Dict
import torch
from datasets import load_dataset
from loguru import logger
from peft import LoraConfig, TaskType, get_peft_model, PeftModel, prepare_model_for_int8_training
from transformers import (
AutoConfig,
BloomForCausalLM,
AutoModel,
AutoModelForCausalLM,
LlamaTokenizer,
LlamaForCausalLM,
BloomTokenizerFast,
AutoTokenizer,
HfArgumentParser,
Trainer,
TrainingArguments,
set_seed,
BitsAndBytesConfig,
deepspeed,
)
from transformers.deepspeed import is_deepspeed_zero3_enabled
from transformers.trainer import TRAINING_ARGS_NAME
from transformers.trainer_pt_utils import LabelSmoother
MODEL_CLASSES = {
"bloom": (AutoConfig, BloomForCausalLM, BloomTokenizerFast),
"chatglm": (AutoConfig, AutoModel, AutoTokenizer),
"llama": (AutoConfig, LlamaForCausalLM, LlamaTokenizer),
"baichuan": (AutoConfig, AutoModelForCausalLM, AutoTokenizer),
"auto": (AutoConfig, AutoModelForCausalLM, AutoTokenizer),
}
IGNORE_INDEX = LabelSmoother.ignore_index
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to fine-tune, or train from scratch.
"""
model_type: str = field(
default=None,
metadata={"help": "Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys())}
)
model_name_or_path: Optional[str] = field(
default=None,
metadata={
"help": (
"The model checkpoint for weights initialization.Don't set if you want to train a model from scratch."
)
},
)
tokenizer_name_or_path: Optional[str] = field(
default=None,
metadata={
"help": (
"The tokenizer for weights initialization.Don't set if you want to train a model from scratch."
)
},
)
model_max_length: Optional[int] = field(default=512, metadata={"help": "The maximum length of the model"})
load_in_8bit: bool = field(default=False, metadata={"help": "Whether to load the model in 8bit mode or not."})
cache_dir: Optional[str] = field(
default=None,
metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
)
use_fast_tokenizer: bool = field(
default=False,
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
)
torch_dtype: Optional[str] = field(
default="float16",
metadata={
"help": (
"Override the default `torch.dtype` and load the model under this dtype. If `auto` is passed, the "
"dtype will be automatically derived from the model's weights."
),
"choices": ["auto", "bfloat16", "float16", "float32"],
},
)
device_map: Optional[str] = field(
default="auto",
metadata={"help": "Device to map model to. If `auto` is passed, the device will be selected automatically. "},
)
trust_remote_code: bool = field(
default=True,
metadata={"help": "Whether to trust remote code when loading a model from a remote checkpoint."},
)
def __post_init__(self):
if self.model_type is None:
raise ValueError(
"You must specify a valid model_type to run training. Available model types are " + ", ".join(
MODEL_CLASSES.keys()))
if self.model_name_or_path is None:
raise ValueError("You must specify a valid model_name_or_path to run training.")
if self.model_max_length < 256:
raise ValueError("You must specify a valid model_max_length >= 256 to run training.")
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
"""
dataset_name: Optional[str] = field(
default=None, metadata={"help": "The name of the dataset to use (via the datasets library)."}
)
dataset_config_name: Optional[str] = field(
default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
)
train_file_dir: Optional[str] = field(default=None, metadata={"help": "The train jsonl data file folder."})
validation_file_dir: Optional[str] = field(default=None, metadata={"help": "The evaluation jsonl file folder."})
max_train_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of training examples to this "
"value if set."
)
},
)
max_eval_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of evaluation examples to this "
"value if set."
)
},
)
overwrite_cache: bool = field(
default=False, metadata={"help": "Overwrite the cached training and evaluation sets"}
)
validation_split_percentage: Optional[int] = field(
default=1,
metadata={
"help": "The percentage of the train set used as validation set in case there's no validation split"
},
)
preprocessing_num_workers: Optional[int] = field(
default=None,
metadata={"help": "The number of processes to use for the preprocessing."},
)
@dataclass
class PeftArguments(TrainingArguments):
use_peft: bool = field(default=True, metadata={"help": "Whether to use peft"})
target_modules: Optional[str] = field(default="all")
lora_rank: Optional[int] = field(default=8)
lora_dropout: Optional[float] = field(default=0.05)
lora_alpha: Optional[float] = field(default=32.0)
modules_to_save: Optional[str] = field(default=None)
peft_path: Optional[str] = field(default=None, metadata={"help": "The path to the peft model"})
qlora: bool = field(default=False, metadata={"help": "Whether to use qlora"})
class CastOutputToFloat(torch.nn.Sequential):
"""Cast the output of the model to float"""
def forward(self, x):
return super().forward(x).to(torch.float32)
class SeparatorStyle:
ADD_COLON_SINGLE = 1
ADD_COLON_TWO = 2
NO_COLON_SINGLE = 3
NO_COLON_TWO = 4
LLAMA2 = 5
CHATGLM = 6
CHATML = 7
CHATINTERN = 8
PHOENIX = 9
@dataclass
class Conversation:
"""A class that manages prompt templates and keeps all conversation history."""
# The name of this template
name: str
# The system prompt
system: str
# Two roles
roles: Sequence[str]
# All messages. Each item is (role, message).
messages: List[List[str]]
# The number of few shot examples
offset: int
# Separators
sep_style: int
sep: str
sep2: str = None
# Stop criteria (the default one is EOS token)
stop_str: str = None
# Stops generation if meeting any token in this list
stop_token_ids: List[int] = None
def get_prompt(self) -> str:
"""Get the prompt for generation."""
if self.sep_style == SeparatorStyle.ADD_COLON_SINGLE:
ret = self.system + self.sep
for role, message in self.messages:
if message:
ret += role + ": " + message + self.sep
else:
ret += role + ":"
return ret
elif self.sep_style == SeparatorStyle.ADD_COLON_TWO:
seps = [self.sep, self.sep2]
ret = self.system + seps[0]
for i, (role, message) in enumerate(self.messages):
if message:
ret += role + ": " + message + seps[i % 2]
else:
ret += role + ":"
return ret
elif self.sep_style == SeparatorStyle.NO_COLON_SINGLE:
ret = self.system
for role, message in self.messages:
if message:
ret += role + message + self.sep
else:
ret += role
return ret
elif self.sep_style == SeparatorStyle.NO_COLON_TWO:
seps = [self.sep, self.sep2]
ret = self.system
for i, (role, message) in enumerate(self.messages):
if message:
ret += role + message + seps[i % 2]
else:
ret += role
return ret
elif self.sep_style == SeparatorStyle.LLAMA2:
seps = [self.sep, self.sep2]
ret = ""
for i, (role, message) in enumerate(self.messages):
if message:
if i == 0:
ret += self.system + message
else:
ret += role + " " + message + seps[i % 2]
else:
ret += role
return ret
elif self.sep_style == SeparatorStyle.CHATGLM:
seps = [self.sep, self.sep2]
if self.system:
ret = self.system + seps[0]
else:
ret = ""
for i, (role, message) in enumerate(self.messages):
if message:
ret += role + ": " + message + seps[i % 2]
else:
ret += role + ":"
return ret
elif self.sep_style == SeparatorStyle.CHATML:
ret = "" if self.system == "" else self.system + self.sep + "\n"
for role, message in self.messages:
if message:
ret += role + "\n" + message + self.sep + "\n"
else:
ret += role + "\n"
return ret
elif self.sep_style == SeparatorStyle.CHATINTERN:
# source: https://huggingface.co/internlm/internlm-chat-7b-8k/blob/bd546fa984b4b0b86958f56bf37f94aa75ab8831/modeling_internlm.py#L771
seps = [self.sep, self.sep2]
ret = self.system
for i, (role, message) in enumerate(self.messages):
if i % 2 == 0:
ret += "<s>"
if message:
ret += role + ":" + message + seps[i % 2] + "\n"
else:
ret += role + ":"
return ret
elif self.sep_style == SeparatorStyle.PHOENIX:
ret = self.system
for role, message in self.messages:
if message:
ret += role + ": " + "<s>" + message + "</s>"
else:
ret += role + ": " + "<s>"
return ret
else:
raise ValueError(f"Invalid style: {self.sep_style}")
def append_message(self, role: str, message: str):
"""Append a new message."""
self.messages.append([role, message])
def to_gradio_chatbot(self):
"""Convert the conversation to gradio chatbot format."""
ret = []
for i, (role, msg) in enumerate(self.messages[self.offset:]):
if i % 2 == 0:
ret.append([msg, None])
else:
ret[-1][-1] = msg
return ret
def copy(self):
return Conversation(
name=self.name,
system=self.system,
roles=self.roles,
messages=[[x, y] for x, y in self.messages],
offset=self.offset,
sep_style=self.sep_style,
sep=self.sep,
sep2=self.sep2,
stop_str=self.stop_str,
stop_token_ids=self.stop_token_ids,
)
def dict(self):
return {
"template_name": self.name,
"system": self.system,
"roles": self.roles,
"messages": self.messages,
"offset": self.offset,
}
# A global registry for all conversation templates
conv_templates: Dict[str, Conversation] = {}
def register_conv_template(template: Conversation, override: bool = False):
"""Register a new conversation template."""
if not override:
assert (
template.name not in conv_templates
), f"{template.name} has been registered."
conv_templates[template.name] = template
# A template similar to the "one_shot" template above but remove the example.
register_conv_template(
Conversation(
name="zero_shot",
system="A chat between a curious human and an artificial intelligence assistant. "
"The assistant gives helpful, detailed, and polite answers to the human's questions.",
roles=("Human", "Assistant"),
messages=[],
offset=0,
sep_style=SeparatorStyle.ADD_COLON_SINGLE,
sep="\n### ",
stop_str="###",
)
)
# Vicuna v1.1 template
register_conv_template(
Conversation(
name="vicuna",
system="A chat between a curious user and an artificial intelligence assistant. "
"The assistant gives helpful, detailed, and polite answers to the user's questions.",
roles=("USER", "ASSISTANT"),
messages=[],
offset=0,
sep_style=SeparatorStyle.ADD_COLON_TWO,
sep=" ",
sep2="</s>",
)
)
# Alpaca default template
register_conv_template(
Conversation(
name="alpaca",
system="Below is an instruction that describes a task. Write a response that appropriately completes the request.",
roles=("### Instruction", "### Response"),
messages=[],
offset=0,
sep_style=SeparatorStyle.ADD_COLON_TWO,
sep="\n\n",
sep2="</s>",
)
)
# ChatGLM default template
register_conv_template(
Conversation(
name="chatglm",
system="",
roles=("问", "答"),
messages=[],
offset=0,
sep_style=SeparatorStyle.CHATGLM,
sep="\n",
sep2="</s>",
)
)
# ChatGLM2 default template
register_conv_template(
Conversation(
name="chatglm2",
system="",
roles=("问", "答"),
messages=[],
offset=0,
sep_style=SeparatorStyle.CHATGLM,
sep="\n\n",
sep2="</s>",
)
)
# Phoenix default template
register_conv_template(
Conversation(
name="phoenix",
system="A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\n\n",
roles=("Human", "Assistant"),
messages=[],
offset=0,
sep_style=SeparatorStyle.PHOENIX,
sep="</s>",
)
)
# Internlm-chat template
register_conv_template(
Conversation(
name="internlm-chat",
system="A chat between a curious <|User|> and an <|Bot|>. The <|Bot|> gives helpful, detailed, and polite answers to the <|User|>'s questions.\n\n",
roles=("<|User|>", "<|Bot|>"),
messages=[],
offset=0,
sep_style=SeparatorStyle.CHATINTERN,
sep="<eoh>",
sep2="<eoa>",
stop_token_ids=[1, 103028],
stop_str="<|User|>",
)
)
# StarChat template
register_conv_template(
Conversation(
name="starchat",
system="<system>\n",
roles=("<|user|>", "<|assistant|>"),
messages=[],
offset=0,
sep_style=SeparatorStyle.CHATML,
sep="<|end|>",
stop_token_ids=[0, 49155],
stop_str="<|end|>",
)
)
# Baichuan-13B-Chat template
register_conv_template(
# source: https://huggingface.co/baichuan-inc/Baichuan-13B-Chat/blob/f5f47be2adbbdceb784f334d6fa1ca2c73e65097/modeling_baichuan.py#L507
# https://huggingface.co/baichuan-inc/Baichuan-13B-Chat/blob/main/generation_config.json
Conversation(
name="baichuan-chat",
system="",
roles=(" <reserved_102> ", " <reserved_103> "),
messages=[],
offset=0,
sep_style=SeparatorStyle.NO_COLON_TWO,
sep="",
sep2="</s>",
stop_token_ids=[2, 195],
)
)
# llama2 template
# reference: https://github.com/facebookresearch/llama/blob/cfc3fc8c1968d390eb830e65c63865e980873a06/llama/generation.py#L212
register_conv_template(
Conversation(
name="llama-2",
system="<s>[INST] <<SYS>>\nYou are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe. "
"Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. "
"Please ensure that your responses are socially unbiased and positive in nature.\n\n"
"If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. "
"If you don't know the answer to a question, please don't share false information.\n<</SYS>>\n\n",
roles=("[INST]", "[/INST]"),
messages=[],
offset=0,
sep_style=SeparatorStyle.LLAMA2,
sep=" ",
sep2=" </s><s>",
stop_token_ids=[2],
)
)
def get_conv_template(name: str) -> Conversation:
"""Get a conversation template."""
return conv_templates[name].copy()
class SavePeftModelTrainer(Trainer):
"""
Trainer for lora models
"""
def save_model(self, output_dir=None, _internal_call=False):
"""Save the LoRA model."""
os.makedirs(output_dir, exist_ok=True)
if self.args.local_rank in [-1, 0]:
torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))
self.model.save_pretrained(output_dir)
def save_model(output_dir, model, tokenizer, args):
"""Save the model and the tokenizer."""
os.makedirs(output_dir, exist_ok=True)
# Take care of distributed/parallel training
model_to_save = model.module if hasattr(model, "module") else model
if args.local_rank in [-1, 0]:
model_to_save.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
torch.save(args, os.path.join(output_dir, TRAINING_ARGS_NAME))
def print_trainable_parameters(model):
"""
Prints the number of trainable parameters in the model.
"""
trainable_params = 0
all_param = 0
for _, param in model.named_parameters():
all_param += param.numel()
if param.requires_grad:
trainable_params += param.numel()
print(
f"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param}"
)
def find_all_linear_names(peft_model, int4=False, int8=False):
"""Find all linear layer names in the model. reference from qlora paper."""
cls = torch.nn.Linear
if int4 or int8:
import bitsandbytes as bnb
if int4:
cls = bnb.nn.Linear4bit
elif int8:
cls = bnb.nn.Linear8bitLt
lora_module_names = set()
for name, module in peft_model.named_modules():
if isinstance(module, cls):
# last layer is not add to lora_module_names
if 'lm_head' in name:
continue
if 'output_layer' in name:
continue
names = name.split('.')
lora_module_names.add(names[0] if len(names) == 1 else names[-1])
return sorted(lora_module_names)
def main():
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, PeftArguments))
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
logger.warning(f"Model args: {model_args}")
logger.warning(f"Data args: {data_args}")
logger.warning(f"Training args: {training_args}")
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+ f" distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
)
# Set seed before initializing model.
set_seed(training_args.seed)
if not model_args.model_type:
raise ValueError("Please specify a model_type, e.g. llama, chatglm, bloom, etc.")
config_class, model_class, tokenizer_class = MODEL_CLASSES[model_args.model_type]
# Load tokenizer
tokenizer_kwargs = {
"cache_dir": model_args.cache_dir,
"use_fast": model_args.use_fast_tokenizer,
"model_max_length": model_args.model_max_length,
"trust_remote_code": model_args.trust_remote_code,
}
tokenizer_name_or_path = model_args.tokenizer_name_or_path
if not tokenizer_name_or_path:
tokenizer_name_or_path = model_args.model_name_or_path
tokenizer = tokenizer_class.from_pretrained(tokenizer_name_or_path, **tokenizer_kwargs)
# tokenizer.padding_side = "right" # set padding side to the right, equal to label's -100 padding side
if tokenizer.pad_token_id is None:
tokenizer.pad_token_id = 0 # set as the <unk> token
logger.debug(f"Tokenizer: {tokenizer}")
# Get datasets
if data_args.dataset_name is not None:
# Downloading and loading a dataset from the hub.
raw_datasets = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
cache_dir=model_args.cache_dir,
)
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
)
raw_datasets["train"] = load_dataset(
data_args.dataset_name,
data_args.dataset_config_name,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
)
else:
# Loading a dataset from local files.
data_files = {}
if data_args.train_file_dir is not None and os.path.exists(data_args.train_file_dir):
train_data_files = glob(f'{data_args.train_file_dir}/**/*.json', recursive=True) + glob(
f'{data_args.train_file_dir}/**/*.jsonl', recursive=True)
logger.info(f"train files: {train_data_files}")
data_files["train"] = train_data_files
if data_args.validation_file_dir is not None and os.path.exists(data_args.validation_file_dir):
eval_data_files = glob(f'{data_args.validation_file_dir}/**/*.json', recursive=True) + glob(
f'{data_args.validation_file_dir}/**/*.jsonl', recursive=True)
logger.info(f"eval files: {eval_data_files}")
data_files["validation"] = eval_data_files
raw_datasets = load_dataset(
'json',
data_files=data_files,
cache_dir=model_args.cache_dir,
)
# If no validation data is there, validation_split_percentage will be used to divide the dataset.
if "validation" not in raw_datasets.keys():
raw_datasets["validation"] = load_dataset(
'json',
data_files=data_files,
split=f"train[:{data_args.validation_split_percentage}%]",
cache_dir=model_args.cache_dir,
)
raw_datasets["train"] = load_dataset(
'json',
data_files=data_files,
split=f"train[{data_args.validation_split_percentage}%:]",
cache_dir=model_args.cache_dir,
)
logger.info(f"Raw datasets: {raw_datasets}")
def preprocess_function(examples):
"""
Preprocessing the datasets.
part of code modified from https://github.com/lm-sys/FastChat
"""
conv = get_conv_template('vicuna')
roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
# Apply prompt templates
conversations = []
sources = examples.get('conversations', []) or examples.get('items', [])
for i, source in enumerate(sources):
if len(source) < 2:
continue
data_role = source[0].get("from", "")
if data_role not in roles or roles[data_role] != conv.roles[0]:
# Skip the first one if it is not from human
source = source[1:]
if len(source) < 2:
continue
conv.messages = []
for j, sentence in enumerate(source):
data_role = sentence.get("from", "")
if data_role not in roles:
logger.warning(f"unknown role: {data_role}, {i}. (ignored)")
continue
if data_role and data_role in roles:
role = roles.get(data_role, "")
if role == conv.roles[j % 2]:
conv.append_message(role, sentence["value"])
if len(conv.messages) < 2 or len(conv.messages) % 2 != 0:
continue
conversations.append(conv.get_prompt())
# Tokenize conversations
input_ids = tokenizer(
conversations,
return_tensors="pt",
padding="max_length",
max_length=tokenizer.model_max_length,
truncation=True,
add_special_tokens=False
).input_ids
targets = input_ids.clone()
# Mask targets. Only compute loss on the assistant outputs.
sep = conv.sep + conv.roles[1] + ": "
for conversation, target in zip(conversations, targets):
total_len = int(target.ne(tokenizer.pad_token_id).sum())
turns = conversation.split(conv.sep2)
cur_len = 0
for i, turn in enumerate(turns):
if turn == "":
break
turn_len = len(tokenizer(turn, add_special_tokens=False).input_ids) + 1 # 1 is </s> token at the end
parts = turn.split(sep)
if len(parts) != 2:
break
parts[0] += sep
instruction_len = len(tokenizer(parts[0], add_special_tokens=False).input_ids) - 1
# Ignore the user instructions
target[cur_len: cur_len + instruction_len] = IGNORE_INDEX
cur_len += turn_len
target[cur_len:] = IGNORE_INDEX
if cur_len < tokenizer.model_max_length:
if cur_len != total_len:
target[:] = IGNORE_INDEX
logger.warning(f"tokenization mismatch: {cur_len} vs. {total_len}. (ignored)")
return dict(
input_ids=input_ids,
labels=targets,
attention_mask=input_ids.ne(tokenizer.pad_token_id),
)
def filter_empty_labels(example):
"""Remove empty labels dataset."""
return not all(label == IGNORE_INDEX for label in example["labels"])
train_dataset = None
max_train_samples = 0
if training_args.do_train:
if "train" not in raw_datasets:
raise ValueError("--do_train requires a train dataset")
train_dataset = raw_datasets['train']
max_train_samples = len(train_dataset)
if data_args.max_train_samples is not None and data_args.max_train_samples > 0:
max_train_samples = min(len(train_dataset), data_args.max_train_samples)
train_dataset = train_dataset.select(range(max_train_samples))
logger.debug(f"Example train_dataset[0]: {train_dataset[0]}")
with training_args.main_process_first(desc="Train dataset tokenization"):
train_dataset = train_dataset.shuffle().map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=train_dataset.column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on dataset",
)
train_dataset = train_dataset.filter(filter_empty_labels, num_proc=data_args.preprocessing_num_workers)
logger.debug(f"Num train_samples: {len(train_dataset)}")
logger.debug("Tokenized training example:")
logger.debug(tokenizer.decode(train_dataset[0]['input_ids']))
eval_dataset = None
max_eval_samples = 0
if training_args.do_eval:
with training_args.main_process_first(desc="Eval dataset tokenization"):
if "validation" not in raw_datasets:
raise ValueError("--do_eval requires a validation dataset")
eval_dataset = raw_datasets["validation"]
max_eval_samples = len(eval_dataset)
if data_args.max_eval_samples is not None and data_args.max_eval_samples > 0:
max_eval_samples = min(len(eval_dataset), data_args.max_eval_samples)
eval_dataset = eval_dataset.select(range(max_eval_samples))
logger.debug(f"Example eval_dataset[0]: {eval_dataset[0]}")
eval_dataset = eval_dataset.map(
preprocess_function,
batched=True,
num_proc=data_args.preprocessing_num_workers,
remove_columns=eval_dataset.column_names,
load_from_cache_file=not data_args.overwrite_cache,
desc="Running tokenizer on dataset",
)
eval_dataset = eval_dataset.filter(filter_empty_labels, num_proc=data_args.preprocessing_num_workers)
logger.debug(f"Num eval_samples: {len(eval_dataset)}")
logger.debug("Tokenized eval example:")
logger.debug(tokenizer.decode(eval_dataset[0]['input_ids']))
# Load model
if model_args.model_name_or_path:
torch_dtype = (
model_args.torch_dtype
if model_args.torch_dtype in ["auto", None]
else getattr(torch, model_args.torch_dtype)
)
world_size = int(os.environ.get("WORLD_SIZE", 1))
ddp = world_size != 1
if ddp:
model_args.device_map = {"": int(os.environ["LOCAL_RANK"]) or 0}
if training_args.qlora and (len(training_args.fsdp) > 0 or deepspeed.is_deepspeed_zero3_enabled()):
logger.warning("FSDP and ZeRO3 are both currently incompatible with QLoRA.")
config = config_class.from_pretrained(
model_args.model_name_or_path,
trust_remote_code=model_args.trust_remote_code,
torch_dtype=torch_dtype,
cache_dir=model_args.cache_dir
)
model = model_class.from_pretrained(
model_args.model_name_or_path,
config=config,
load_in_8bit=model_args.load_in_8bit,
low_cpu_mem_usage=(not is_deepspeed_zero3_enabled()),
device_map=model_args.device_map,
trust_remote_code=model_args.trust_remote_code,
quantization_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch_dtype,
) if training_args.qlora else None,
)
if hasattr(model, 'lm_head'):
model.lm_head = CastOutputToFloat(model.lm_head)
else:
raise ValueError(f"Error, model_name_or_path is None, SFT must be loaded from a pre-trained model")
if training_args.use_peft:
logger.info("Fine-tuning method: LoRA(PEFT)")
if training_args.peft_path is not None:
logger.info(f"Peft from pre-trained model: {training_args.peft_path}")
model = PeftModel.from_pretrained(model, training_args.peft_path, is_trainable=True)
else:
target_modules = training_args.target_modules.split(',') if training_args.target_modules else None
if target_modules and 'all' in target_modules:
target_modules = find_all_linear_names(model, int4=False, int8=model_args.load_in_8bit)
modules_to_save = training_args.modules_to_save
if modules_to_save is not None:
modules_to_save = modules_to_save.split(',')
logger.info(f"Peft target_modules: {target_modules}")
logger.info(f"Peft lora_rank: {training_args.lora_rank}")
peft_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
target_modules=target_modules,
inference_mode=False,
r=training_args.lora_rank,
lora_alpha=training_args.lora_alpha,
lora_dropout=training_args.lora_dropout,
modules_to_save=modules_to_save)
model = get_peft_model(model, peft_config)
if model_args.load_in_8bit:
model = prepare_model_for_int8_training(model)
model.print_trainable_parameters()
else:
logger.info("Fine-tuning method: Full parameters training")
model = model.float()
print_trainable_parameters(model)
logger.debug(f"Model: {model}")
# Initialize our Trainer
if training_args.gradient_checkpointing:
model.gradient_checkpointing_enable()
model.config.use_cache = False
else:
model.config.use_cache = True
model.enable_input_require_grads()
if not ddp and torch.cuda.device_count() > 1:
# Keeps Trainer from trying its own DataParallelism when more than 1 gpu is available
model.is_parallelizable = True
model.model_parallel = True
trainer = SavePeftModelTrainer(
model=model,
args=training_args,
train_dataset=train_dataset if training_args.do_train else None,
eval_dataset=eval_dataset if training_args.do_eval else None,
tokenizer=tokenizer,
)
# Training
if training_args.do_train:
logger.info("*** Train ***")
sample = next(iter(trainer.get_train_dataloader()))
logger.debug(f"Train dataloader example: {sample}")
logger.debug(f"Detail input_ids: {list(sample['input_ids'])[:3]}, \nlabels: {list(sample['labels'])[:3]}")
logger.debug(f"Decode input_ids[0]: {tokenizer.decode(sample['input_ids'][0])}")
replaced_labels = [label if label != IGNORE_INDEX else tokenizer.pad_token_id for label in sample['labels'][0]]
logger.debug(f"Decode labels[0]: {tokenizer.decode(replaced_labels)}")
checkpoint = None
if training_args.resume_from_checkpoint is not None:
checkpoint = training_args.resume_from_checkpoint
train_result = trainer.train(resume_from_checkpoint=checkpoint)
metrics = train_result.metrics
metrics["train_samples"] = max_train_samples
logger.debug(f"Training metrics: {metrics}")
trainer.log_metrics("train", metrics)
trainer.save_metrics("train", metrics)
trainer.save_state()
logger.info(f"Saving model checkpoint to {training_args.output_dir}")
save_model(training_args.output_dir, model, tokenizer, training_args)
# Evaluation
if training_args.do_eval:
logger.info("*** Evaluate ***")
metrics = trainer.evaluate()
metrics["eval_samples"] = max_eval_samples
try:
perplexity = math.exp(metrics["eval_loss"])
except OverflowError:
perplexity = float("inf")
metrics["perplexity"] = perplexity
logger.debug(f"Eval metrics: {metrics}")
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)
if __name__ == "__main__":
main()